Human behaviour modelling for welfare technology using hidden Markov models
Veralia GabrielaSánchez
aOla MariusLysaker
bNils-OlavSkeie
aa
Department of Electrical Engineering, Information Technology and Cybernetics, University of South-Eastern Norway (USN), Porsgrunn, Norway
b
Department of Process, Energy and Environmental Technology, University of South- Eastern Norway (USN), Porsgrunn, Norway
Pattern Recognition Letters, 9.
DOI: http://dx.doi.org10.1016/j.patrec.2019.09.022
This article has been accepted for publication and undergone full peer review but has not been through the copyediting, typesetting, pagination and proofreading process, which may lead to differences
between this version and the Version of Record. This article is protected by copyright. All rights reserved.
ABSTRACT
Human behaviour modelling for welfare technology is the task of recognizing a person's behaviour patterns in order to construct a safe environment for that person. It is useful in building environments for older adults or to help any person in his or her daily life. The aim of this study is to model the behaviour of a person living in a smart house environment in order to detect abnormal behaviour and assist the person if help is needed. Hidden Markov models, location of the person in the house, posture of the person, and time frame rules are implemented using a real-world, open-source dataset for training and testing. The proposed model presented in this study models the normal behaviour of a person and detects anomalies in the usual pattern. The model shows good results in the identification of abnormal behaviour when tested.
The work presented here is based on the theory that people
1. Introduction generate patterns in their daily activities and behaviour [6,7].
Therefore, a repetitive pattern in the person's behaviour helps to The older population in Nordic and other European countries recognize, model, and predict future events.
has substantially increased. In the European Union, 5.5% of the In order to recognize the behaviour of a person, Hidden population is aged 80 or above as of 2017. This number will almost Markov models (HMM) were used. Posture checking and time double to 12.7% by 2080 [1]. In Norway, 38.5% of households frame logic were added as an extra layer of recognition to model with people aged 65 and over live alone [2]. the behaviour of the person as normal or abnormal.
In Nordic countries, the term welfare technology refers to HMM was used because it is a statistical method that assumes
“technology used for environmental control, safety and well-being a Markov process with missing or unobserved states. Moreover, in particular for elderly and disabled people” [3]. Welfare the Markov assumption is a sequence of events in which the technology is more often referred to as ambient assisted living probability of each event depends only on the previous event.
outside of Scandinavia. For the modelling, a real world, open-source dataset was used.
In this work, human behaviour modelling (HBM) is proposed The dataset comprised a finite number of days, in which half the as a type of welfare technology that can recognize an individual’s days were used for training and the other half for testing. Once the behaviour patterns in a smart house, thereby helping to construct a HMM model was trained, the Viterbi algorithm was used to test safe environment. Smart house development is important for those the validity of the model for the remaining days in the dataset. The who prefer to live in their own homes as long as they can care for Viterbi algorithm enables detecting whether the input sequence for
themselves, and are defined as living environments designed to testing (or a new input sequence) is classified as a normal or an assist residents with their daily activities and to promote abnormal behaviour.
independent lifestyles [4].
Therefore, HBM is developed to detect abnormal behaviour 1.1 Aim and Objectives
(anomalies) in a person's behaviour patterns and provide assistance The aim of the present study is to model the behaviour of a if needed. “Anomaly detection refers to the problem of finding person living in a smart house environment, to detect abnormal patterns in data that do not conform to expected behaviour”[5]. behaviour and alert family or a caretaker if assistance is needed.
Examples of abnormal behaviour could be falls and early signs of The main objective of using HMM for human behaviour
dementia. modelling is to predict whether the current activity is normal or
In the present study, the term behaviour refers to the activity, abnormal. HMM is used because it is a statistical method that duration of activity, location, and posture of a person. Recognizing works well with a small dataset or insufficient training data [8,9].
a person's behaviour patterns helps in constructing a safer environment for older adults, people with disabilities, and a more comfortable lifestyle in general for any person.
Activity: ADL.
1.2 Activity vs Behaviour combination with other methods such as neural network and In the present study, the term activity refers to the actions of
people in a specific room or area. These actions are known as activities of daily living (ADL) and include sleeping, personal hygiene, showering, dressing, undressing, eating, etc. ADLs are more formally defined as the actions that require “basic skills and focus on activities to take care of one's own body”.
The term behaviour in the present study refers to the combination of activity, duration, location and posture of the person.
Location: place in the house where the person is doing the activity (bedroom, bathroom, kitchen, etc.).
Posture: position of the person (lying, sitting, standing).
Duration: The time span from the start to the end of an activity, given in hours, minutes and seconds (hh:mm:ss).
An example of a behaviour can be having breakfast. The breakfast behaviour comprises eating (activity), being in the kitchen (location), sitting (posture) and within a time span (duration). Behaviour can also be a sequence of activities.
The rest of this paper is organized as follows. In section 2, presents the background. Section 3 discusses the methods. Section 4 describes the experiments. Section 5 presents the results and section 6 discusses the results. Finally, the conclusion is provided in section 7.
intelligent agents [22,23]. The center for advanced studies in adaptive systems (CASAS), from Washington State University, implemented HMM with promising results with several residences in a smart house [24]. Another study on HAR for diabetes patients in a smart house was developed using HMM with 98% accuracy [25].
Several other studies are not traditionally called HAR but ADL recognition or detection. Both HAR and ADL recognition works have the same aim. A two-stage, multi-Markov model for ADL detection was used by Kalra et al. [26]. Another study used the dataset from CASAS to recognize ADL [27] and compared support vector machines (SVM), HMM, Fishe kernel learning (FKL) and random forests (RF). Gayathri et al. [28] also used the dataset from CASAS to detect ADL using Markov logic networks.
In addition, ADL recognition has been studied with HSMM.
One of the first research on modelling ADL is the work of Duong et al. [10], who used a Switching Hidden Semi-Markov model (S- HSMM) for activity recognition and abnormal detection in a pervasive environment. Duong created a double-layered extension of HSMM. That model focused on distinguishing a person's major routines (making breakfast, eating breakfast, etc.). Another very important work on HSMM for HAR using on real world activity recognition data is the work of Van Kasteren et al. [29].
The studies mentioned so far focus on HAR/ADL recognition using HMM. There are, however, other studies that use other techniques for HAR recognition, such as machine learning techniques.
2. Background One machine learning technique for anomaly detection is
Improving the lives and safety of older adults has been an important area of research with regard to smart house welfare technology [10–12]. This generally includes detecting falls, among other issues, and warning family or caretakers of any potential dangerous or abnormal situation [13–15].
Ideally, a smart house designed to help people should search for patterns in the user's activity or behaviour and detect any deviation from this pattern. Other projects aim to ease a person's daily life regardless of age, while increasing comfort and security [16].
The general technique used to achieve the aforementioned goals is Human Activity Recognition (HAR), which is the task of recognizing the activities of a person. There are several analysis methods used for HAR[17]. However, studies involving human behaviour modelling have not received the same kind of attention.
One study involving HAR modelling used decision trees with promising results of 88.02% for the activity recognition task.
However, they study concluded that for modelling human activities, decision trees did not meet the expectations [18].
Therefore, in this section, we use the knowledge derived from the state-of-the-art techniques in HAR as a foundation for the present work in human behaviour modelling. Among the most popular analysis methods used for HAR are machine learning techniques and statistical methods.
Hierarchical Temporal Memory (HTM). Some studies have used HTM for anomaly detection in streaming data, online sequence learning and short-term forecasting of electrical load time series [30–32]. However, there are very few works using HTM for abnormal behaviour detection for welfare technology.
Although machine learning techniques could be useful for HBM, they require large amounts of data [19]. This is particularly a problem in smart house environments, where it is difficult to obtain large relevant datasets. Finding a person who is willing to live in a smart house that is set up to collect data is challenging, especially if the final user in an older person, where many ethical challenges are involved [4]. Therefore, the data used in this study is an open-source small dataset.
In addition, anomalies in the behaviour of a person cannot as of yet be learned from the currently available dataset. The publicly available datasets comprise a person living alone for a few days or months and performing his or her daily activities. The datasets do not contain any abnormal behaviour in their patterns. Therefore, in the present study, we have created a fictional dataset with abnormal behaviour to test our model.
The studies described in this section show that HMM has been implemented with good results for HAR/ADL. Therefore, in the present study, HMM is implemented for the first step of human behaviour modelling.
Machine learning techniques require a large amount of data for training. In contrast, classic statistical methods are more effective
than machine learning techniques when a smaller dataset is used 3. Methods [19]. Therefore, for pattern recognition within smart houses, a
statistical approach tends to perform better since the datasets are 3.1 Design
usually small. First, the HMM algorithm was implemented. The activities of
Useful statistical algorithms include HMM and Hidden Semi- the person can be modelled as Markov chains. Therefore, HMM Markov Models (HSMM). HMM have been used for several other was implemented in order to train the algorithm and recognize the tasks with excellent results, such as speech recognition, pattern activities. The person's activities were used as the hidden states, recognition and artificial intelligence [20]. and the observable states were the person's location (obtained from
One of the first works on HMM for HAR is the work of the sensor data in the dataset). MATLAB was used to implement Yamato et al. [21]. Later on, HMM has been used separately or in the HMM.
Table 1:Attributes of the OrdonezA dataset used [36]
Name Value
Setting Apartment
Number of locations 4 Rooms and hall/entrance
Number of labelled days 14 days
Labels (ADLs included) Leaving, Toileting (Personal
hygiene),Showering, Sleeping, Breakfast, Lunch, Dinner, Snack, Spare Time/TV, Grooming Number of sensors 12 sensors
Sensors PIR: Shower, Basin, Cooktop
Magnetic: Maindoor, Fridge, Cabinet, Cupboard
Fig. 1 Flow diagram of proposed methodology for HBM Flush: Toilet
Pressure: Seat, Bed
The HMM was trained using the Baum-Welch algorithm. Electric: Microwave, Toaster
Later, the Viterbi algorithm was used to find the most probable activity given the input sequence.
The HMM module outputted the activity of the person. The person's posture was then checked using a Boolean method to detect abnormal behaviour such as falls (section 4.2). Afterwards, time rules were applied, considering the duration for each activity.
The time frame rules enable more accurate detection of any abnormal behaviour in the person (section 4.3). Fig. 1 shows the proposed methodology used in the present work.
HSMM was also used in the present study as an alternative to HMM. HSMM was chosen because HSMM “can have any arbitrary duration distribution” [33] . The HSMM implementation was programmed with the mhsmm package for R [34]. The main reason for using HSMM was to compare the performance of HMM
Note that in the present work, we used the same terms as they are used in the dataset. Therefore, the term toileting is used as it is instead of the more proper term personal hygiene.
There are two instances of data, one of 14 days (OrdonezA), and one of 21 days (OrdonezB), corresponding to each user, person A and person B, respectively. The activities were manually labelled by the users. Both dataset OrdonezA and dataset OrdonezB were used in this study.
An open-source dataset was chosen in order to obtain unbiased results. In addition, this dataset comprises real-world data and it has been used in other research [36]. The fourteen days' activity data for the first dataset is shown in Fig. 2.
and HSMM with regard to HBM.
3.2 Data 3.2.1 Data Handling
The dataset is a text file that can be imported into MATLAB.
The data used for this study come from an open dataset named Each day in the dataset represents a day sequence.
Activities of Daily Living (ADLs) Recognition Using Binary Sensors Data Set, available for download [35]. The data collected comprise information about the ADLs gathered by two people living on a daily basis in their own homes. The dataset is in a text file format.
The dataset properties are depicted in Table 1. The information in the dataset includes the date, time, location of the sensor, type of sensor, location (room) in the apartment and activities.
For OrdonezA, seven days were used as a training sequence, and another seven days were used as a testing sequence. For OrdonezB, fourteen days were used as a training sequence, and another seven days were used as a testing sequence.
The house room location was chosen as the main observed variable, and the activities were the hidden data. Table 2 depicts the first day from the dataset.
Fig. 2 Activity Data Graph. Each day is shown in a different colour, according to the legend at the right. The first activity is always sleeping and the last activity is leaving. The activities are shown in a sequence.
Standing 3
Sitting 2
Lying 1
Posture Number
3.
2.
1.
The initial state distribution (π) The emission probability (B) The transition probability (A)
Sleeping
10:18:11 02:27:59
28- -
Activity End Time
Start Time Date
11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11
Table 2: Example of Day 1 in the dataset. The dataset comprises the variable (sometimes called the event) occurs at each instant of date, start time, end time, activity and location. To train the HMM,
location was used as an observable variable and activity was used as the hidden variable
Location Bedroom
time. The hidden variables, Ys, can be observed by another set of stochastic processes that produce the sequence of observations Xs [9]. Each hidden state y can emit one and only one possible observed variable x.
28- - 10:21:24 10:23:36 Toileting Bathroom The HMM is fully determined by three probabilities:
28- - 28- - 28- - 28- - 28- - 28- - 28- - 28- - 28- - 28- - 28- - 28- -
10:25:44 10:34:23 10:49:48 10:51:41 13:06:04 13:09:31 13:38:40 14:22:38 14:27:11 15:04:59 15:07:01 20:20:55
10:33:00 10:43:00 10:51:13 13:05:07 13:06:31 13:29:09 14:21:40 14:27:07 15:04:00 15:06:29 20:20:00 20:20:59
Showering Breakfast Grooming Spare_Time/TV Toileting Leaving Spare_Time/TV Toileting Lunch Grooming Spare_Time/TV Snack
Bathroom Kitchen Bathroom Living Room Bathroom Hall Living Room Bathroom Kitchen Bathroom Living Room Kitchen
1.
p(yt+1|yt) 2.
p( xt| yt) 3.
π(y0)
The aim of the HMM is to solve the following three problems:
Evaluation: Inferring the probability of an observation sequence given the fully characterized model.
Decoding: Finding the path of hidden states that most probably generated the observed output.
Learning: Optimizing the model that best describes how 28- - 20:21:15 02:06:00 Spare_Time/TV Living Room a given sequence of observations (also known as a
training sequence) occurs.
An additional column for posture was added to the dataset. The In the present study, the focus is on learning and decoding to values of the posture variable were lying, sitting and standing. The recognize the behaviour of the person.
values were assigned according to each activity and coded to
numbers in order to develop the MATLAB code. Table 3 shows 3.3.1 Hidden Semi-Markov Model
the values allowed for each activity. HMM does not take into account the duration parameter of the A total of 10 activities, 5 locations and 3 postures were used. current activity. Therefore, to model the behaviour of the person, Table 3: Assigned number and posture of the person according to the
activities
Allowed activity Sleeping, Spare Time
Personal Hygiene, Breakfast, Spare Time, Snack, Lunch, Dinner
Showering, Grooming, Leaving
an additional layer is needed to consider the duration.
To overcome this additional layer, HSMM were also studied as an alternative solution for behaviour modelling.
HSMM is an extension of HMM. HSMM allows “the underlying process to be a semi-Markov chain with a variable duration or sojourn time for each state” [37].
The HSMM can produce a sequence of observations. The number of observations that are emitted during state i is constrained by the length of time spent (duration) in state i, usually 3.3 Hidden Markov Model represented as d. Thus, for each state i, there is a specified duration
distribution Di, which can be parametric or non-parametric.
As a result, the HSMM parameters are the same as the HMM, plus the sojourn time for each state. λ= (A,B,D,π).
Fig. 4 shows a representation of the HSMM. The HSMM runs from 1 to T times, where x is the current observation, y is the current state and d is the duration variable of the current state. Thus at each step, the variable comprises Vt={yt,d,xt}. Each state duration can be modelled by any distribution in the exponential family.
Fig. 3: Schematics of HMM representation
The idea of Hidden Markov Models (HMM) was first introduced in the late 1960s [9]. HMMs are a “subclass of Bayesian networks known as dynamic Bayesian networks” [20].
HMM is a generative probabilistic model that is used for generating hidden states from observable data.
HMMs utilize Bayesian rules such that a separate model p(x|y) is learned for each class. Therefore, the posterior probability p(y|x)
can be calculated. Fig. 4: Schematics of HSMM representation
Fig. 3 shows a schematic of HMM. The Xs are the observable variables and the Ys are the hidden variables. The observed
is decoding task.
𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑𝑆𝑡𝑎𝑡𝑒 Fig. 5: Location transition with heat transition scale probability on the
right side. The transitions were estimated using the Baum-Welch algorithm. Fig. 6: Activity transition probabilities after the HMM was trained using the Baum-Welch algorithm
sequence of hidden states that produced that observable sequence.
4. Experiments The Viterbi algorithm was used to achieve th
The experiments were performed for both available datasets, OrdonezA and OrdonezB. However, in this section, the figures and
The observation sequence was the location data of a day. A total of seven days was used for testing, comprising the last seven days of each dataset.
explanations are only provided for the first dataset OdonezA in The prediction accuracy was found by computing the order to keep the text as clear as possible. Section 5 provides the
results for both datasets. The experiments with the OrdonezB percentage of the real or actual sequence state that agreed with the predicted sequence. Hence, the formula:
dataset follow the same procedure as described here for the
OrdonezA dataset. ∑ 𝑎𝑐𝑡𝑢𝑎𝑙𝑆𝑡𝑎𝑡𝑒
= 𝑡𝑜𝑡𝑎𝑙𝑆𝑡𝑎𝑡𝑒
4.1 HMM and HSMM 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = ∑
The HMM was built by initializing the transition and emission 4.1.3 HSMM probabilities (A and B). The activities data corresponded to the
hidden state, and the location data corresponded to the observed data (traning sequence). The HSMM was built with the sojourn time D.
The HSMM was trained using the mhsmm package for R. The same parameters and training data as in the HMM were used. Only the duration parameter D was added to model the HSMM.
For the parameter D, the Gamma distribution was first used to 4.1.1 The Learning Task model the duration of each activity. Second, a Poisson distribution The learning process deals with how to adjust the model was also used to model the duration of each activity.
parameters λ= (A,B,π) to maximize the probability of the 4.2 Checking posture to detect possible fall observation sequence p(X|λ), where X represents the observed
sequence, X=x1,x2,x3...xn. The idea is to optimize the model The posture of the person was checked using a Boolean logic parameters that best describe how a given observation sequence is method. The posture variable can have only three values: lying, produced [9]. The Baum-Welch algorithm is used for this. sitting and standing. The primary check was applied to the posture
For the OrdonezA dataset, the first seven days were used to lying because it could indicate a fall.
train the model; this is known as the training data. The last seven A priori knowledge was applied for this step. Therefore, the days were used for testing. For the OrdonezB dataset, the first lying posture was only allowed in the sleeping and spare time fourteen days were used to train the model and the last seven days activities, as shown in table Table 3. There should not have been
were used for testing. any lying in any activity performed in the locations: bathroom,
The training sequence uses the location data as the observable kitchen and entrance. If the posture lying was found in any of the sequence. The dataset comprises five room locations, as depicted aforementioned locations, then a fall was detected.
in Table 1.
Fig. shows the transition between the locations, estimated 4.3 Time frame rules
with the Baum-Welch algorithm. That is, A(i,j), the probability of Time frame rules were used to determine whether the current transition from state i to state j, given the training input sequence activity had a reasonable duration. The seven days that were used
of location. for training the HMM were also used for training the duration of The hidden variables were the activities. A total of ten each activity.
activities were trained. However, there was no data for the activity First, the duration of each activity was extracted from the dinner in the training dataset OrdonezA. Fig. shows the trained dataset. Then, the minimum and maximum duration for each HMM corresponding to the transition between the activities. In activity in the training dataset was computed. Finally, to calculate addition, a probability transition heat map of the activities is given the duration range for each activity, a ±20% threshold was used.
in supplementary material (Fig. A.1). Therefore, 20% was added to the maximum duration of each 4.1.2 The Decoding: Viterbi Algorithm
Once the model was trained and given a new observation
activity, and 20% was deducted from the minimum duration of each activity.
sequence, it was possible to determine the best, most likely
patterns.
its in behaviour abnormal
about information
on for this
hows Fig. 7: Results for the OrdonezA dataset for Day 8 and Day 12: Blue circles show the real data. Red crosses show the estimated results
Maximum Duration Allowed = MaximumDuration + 20% is true for most of the other days. One possible reas
Minimum Duration Allowed = MinimumDuration − 20% estimation is that both of these behaviours are in the bathroom If the person spent more than the maximum or less than the location.
minimum duration allowed for any activity, an abnormal In addition, the behaviour breakfast was predicted instead of behaviour alert was triggered. the correct lunch and snacking behaviours. The same reason as
4.4 Fictional dataset test before could be applied here, that the behaviours breakfast, lunch and snacking are all located in the kitchen. Therefore, the The dataset used in the present study did not contain
As mentioned in section 2, we were unable to find a relevant dataset containing both normal and abnormal behaviour for a person living in a smart house. Therefore, it was not possible to learn abnormal behaviour from the dataset used in the present project.
To overcome this challenge, we manually created a small fictional dataset containing some abnormal behaviour comprising of three days. The data in the dataset contained some changes in the duration, posture, and sequence of activities. These three fictional days were created to test whether the model presented in this work could detect abnormal behaviour in the person's behaviour.
Therefore, the duration of some of the activities was exaggerated. In addition, on the first fictional day, the posture for
prediction accuracy was 72% using the HMM algorithm.
After the HMM results were obtained, the posture of the person was checked. There were no warnings when the posture was checked on any of the days assigned for testing because there was no abnormal behaviour in the testing dataset.
Finally, time rule was applied to the model to check the duration. Table 4 shows the warnings in the behaviour when the duration was checked. The results show that the model can detect whether the user has spent either too much or too little time performing a behaviour.
Most of the warnings are for the behaviour breakfast, because the duration of that behaviour varied a great deal. One reason for this is that the behaviour breakfast was predicted instead of the real behaviour snack or lunch, which usually take less time (snack) or more time (lunch).
the activity leaving at the Entrance was changed to lying, to Table 4: Warnings in behaviours when duration is checked for the simulate a fall. Finally, the sequence of activities on the third day OrdonezA dataset. The first column is the day in the testing data. The
was changed. second column shows the predicted behaviour. The third column s
the duration of the predicted behaviour.
Day Behaviour Duration MinTime MaxTime
5. Results Day 8
Day 8
Grooming Sleeping
00:00:05 00:00:04
00:00:09 10:13:45
00:11:16 10:09:18
5.1 Results on first dataset OrdonezA Day 8 Leaving 00:16:59 00:18:28 03:30:26
Day 8 Breakfast 00:43:08 00:03:40 00:12:44
As described in section 4.1.2, the Viterbi algorithm estimates the behaviour of the person based on the observation sequence of the observable variable location. The last seven days were used for testing. The predicted behaviours were compared to the actual behaviours sequence in the dataset. The results for the OrdonezA dataset Day 8 and Day 10 are shown in Fig. . The supplementary
material conta ins the results for all the test days in Ordone zA
d a t
a s et ( F i
g . A . 2 )
.
For the first day in the tes ting dataset, i t is possible t o see that
the predicted beh aviour is groomi ng instead of toil eting. The same
Day 9 Day 9 Day 10 Day 10 Day 11 Day 11 Day 13 Day 14 Day 14
Grooming Breakfast Grooming Breakfast Leaving Grooming Leaving Grooming Breakfast
00:13:38 00:35:05 00:13:41 00:35:56 03:49:40 00:00:02 04:03:00 00:15:46 00:52:05
00:00:09 00:03:40 00:00:09 00:03:40 00:18:28 00:00:09 00:18:28 00:00:09 00:03:40
00:11:16 00:12:44 00:11:16 00:12:44 03:30:26 00:11:16 03:30:26 00:11:16 00:12:44
behaviour e
activity
Fig. 8: Results for the three days in the fictional dataset. Legend in x axis is as follow: 1) Breakfast, 2) Grooming, 3) Leaving, 4) Lunch, 5) Showering, 6) Sleeping, 7) Spare Time/TV, 8) Toileting
Other warnings resulted for the behaviour grooming and behaviours breakfast, lunch and snack all take place in the location leaving. Hence, the model detected that the person spent less time kitchen.
or more time than usual engaging in these behaviours and issued To overcome this challenge in the prediction, the ideal dataset
warnings. should state the activity eating instead of the behaviour breakfast.
In order to test whether the model detected abnormal However, a limitation was that the dataset used in the present study behaviour, a fictional dataset as described in section 4.4 was also is a combination of activity and behaviour. Therefore, breakfast is
used for testing. a behaviour that includes the activity eating. The difference
The results for the fictional dataset are shown in Fig. 8. The between behaviour and activity is defined in section 1.2. This results for the three fictional days show the behaviour grooming means that the model presented in the current study was able to predicted instead of the behaviour toileting or showering. In detect that the person was eating in the kitchen, but it could not addition, the second fictional day shows the prediction of the detect whether the person was having breakfast, lunch or snacking.
behaviour breakfast instead of lunch. In addition, there should be enough training data to model a Regarding the check on the posture of the person, a warning person's normal and abnormal behaviours. It is worth noting that was issued for the first fictional day: “Warning: Person lying in the method used in the study is a probabilistic model. As a result, Entrance, possible fall detected”. Therefore, the model the behaviour predictions are based on the highest probable path successfully checked the posture of the person and abnormal given an input sequence. Hence, HMM has been shown to be a
behaviour such as a fall was detected. good method for HBM.
The model also issued another warning for the second fictional HSMM was also studied for modelling the behaviour because day as follows: “Warning: Person is not lying while in bed”. Thus, the duration is modelled explicitly. However, for the aim of this the model detected the abnormal behaviour in the behaviour study, HSMM does not meet expectations. The main reason for the
sleeping. poor results obtained with the HSMM is that a person's behaviour
Other warnings were issued by the model regarding the must follow the exact same pattern all the time in order to model duration of some of the behaviour. The model detected that the the duration.
person spent less time than usual engaging in the behaviour Previous work in HSMM for activity recognition has leaving for the second and third fictional days. Therefore, a effectively distinguished between having breakfast, lunch or warning was generated, stating that “The person came back too dinner [33]. However, those experiments were constrained to the
early”. person following the same pattern of opening the fridge and then
5.2 Results on the second dataset OrdonezB using the stove, sink, cupboard and table in the same sequence.
Only the duration changed for the sequence.
The result graphs for the OrdonezB dataset are in the supplementary material (Fig. A.3). In short, fourteen days were used for training and seven days were used for testing.
The model predicted the behaviour grooming for several of the days in the testing dataset, instead of the real behaviour toileting
or showering. As with the OrdonezA dataset results, this prediction was made because all three behaviours are performed in the bathroom location.
Similarly, the model predicted the behaviour breakfast instead of lunch or snack for several days in the testing dataset. The same reason as before applies here too: the behaviours breakfast, lunch
and snack are all performed in the same kitchen location. For modelling human behaviour in the present study, the constraint of following the same exact sequence is impractical, since people do not always follow the same strict pattern within each behaviour. Therefore, it was not possible to train the HSMM effectively as the dataset does not contain the same strict pattern for every day. Consequently, HMM was chosen as the best method for modelling a person's behaviour.
When the posture of the person was checked, the results showed that the model was able to detect a fall in the entrance in the dataset that was fictionally created. Hence, Boolean logic is a fast and effective method for the purpose of fall detection.
Lastly, the duration for each behaviour was checked using time frame rules. The current duration of each behaviour was extracted from the dataset. A ± 20% approach was implemented to
6. Discussion determine whether the current behaviour was within the normal
duration. The results showed that the model could effectively In this study, a Hidden Markov model is used for predicting detect whether the person has spent too much or too little time in the behaviour of a person. The accuracy of the HMM is 72%. The an .
results for both dataset OrdonezA and OrdonezB showed The posture was checked before the duration because in our consistency in the predictions. Most of the mispredictions that stimation, detecting a fall is more important than the duration of occurred were for the behaviour breakfast, primarily because the a .
– 148.
Finally, the model in the present study is tested offline. Networking, Sens. Control. 2008. ICNSC 2008. IEEE Int. Conf., However, in the future, the model should be run in real-time. This
means there should be a period of learning. The results presented in this study show that our model was able to learn from seven
[16]
2008: pp. 1505–1510.
T. Van Kasteren, A. Noulas, G. Englebienne, B. Kröse, Accurate activity recognition in a home setting, in: Proc. 10th Int. Conf.
Ubiquitous Comput., 2008: pp. 1–9.
days. In addition, the real-time model should be able to check the [17] V.G. Sanchez, C.F. Pfeiffer, N.-O. Skeie, A Review of Smart behaviour of the person at fixed time intervals in order to detect House Analysis Methods for Assisting Older People Living Alone, whether the behaviour is normal or abnormal, such as using a time
event driven logic every sixty seconds. [18]
J. Sens. Actuator Networks. 6 (2017) 11.
V.G. Sánchez, N.-O. Skeie, Decision Trees for Human Activity Recognition in Smart House Environments, in: Proc. 59th Conf.
Simul. Model. (SIMS 59), 26-28 Sept. 2018, Oslo Metrop. Univ.
Norw., 2018: pp. 222–229.
7. Conclusions and Future Work [19] D. Bzdok, N. Altman, M. Krzywinski, Statistics versus machine learning, Nat. Methods. 15 (2018) 233.
Human behaviour modelling (HBM) for welfare technology is proposed to detect abnormal behaviour. HBM allows detecting any deviation from the usual or normal pattern of the person. Hence,
[20]
[21]
Z. Ghahramani, An introduction to hidden Markov models and Bayesian networks, Int. J. Pattern Recognit. Artif. Intell. 15 (2001) 9–42.
J. Yamato, J. Ohya, K. Ishii, Recognizing human action in time- abnormal behaviours are possible to detect and alert a family sequential images using hidden markov model, in: Comput. Vis.
member or caretaker if the person is in need of any assistance. Pattern Recognition, 1992. Proc. CVPR’92., 1992 IEEE Comput.
The behaviour of the person consists of the activity performed by the person, the duration, the location and posture of the person.
Hidden Markov models (HMM), is used to model and predict the behaviour of the person. The experimental evaluation shows good
[22]
[23]
Soc. Conf., 1992: pp. 379–385.
A. Mihailidis, J.N. Boger, T. Craig, J. Hoey, The COACH prompting system to assist older adults with dementia through handwashing: An efficacy study, BMC Geriatr. 8 (2008) 28.
C. Zhang, W.A. Gruver, Distributed agent system for behavior results using an open source real world dataset. pattern recognition, in: 2010 Int. Conf. Mach. Learn. Cybern.,
Future work should focus on studying other algorithms, including statistical, machine learning and deep learning with the aim of developing HBM with possible improvements and compare them with the performance of the present study.
[24]
[25]
2010: pp. 204–209.
A.S. Crandall, D.J. Cook, Coping with multiple residents in a smart environment, J. Ambient Intell. Smart Environ. 1 (2009) 323–334.
A. Helal, D.J. Cook, M. Schmalz, Smart home-based health platform for behavioral monitoring and alteration of diabetes
patients, J. Diabetes Sci. Technol. 3 (2009) 141
References [26] L. Kalra, X. Zhao, A.J. Soto, E. Milios, Detection of daily living
activities using a two-stage Markov model, J. Ambient Intell. Smart [1] EuropeanCommission, Increase in the share of the population aged Environ. 5 (2013) 273–285.
65 years or over between 2007 and 2017, (2018). [27] C. Debes, A. Merentitis, S. Sukhanov, M. Niessen, N. Frangiadakis, https://ec.europa.eu/eurostat/statistics- A. Bauer, Monitoring activities of daily living in smart homes:
explained/index.php/Population_structure_and_ageing#The_share_ Understanding human behavior, IEEE Signal Process. Mag. 33
of_elderly_people_continues_to_increase. (2016) 81–94.
[2] S. sentralbyraa, Key figures for the population, 2018, (2018). [28] K.S. Gayathri, K.S. Easwarakumar, S. Elias, Probabilistic ontology
www.ssb.no/en/folkfram. based activity recognition in smart homes using Markov Logic
[3] R. Brynn, Universal design and welfare technology, Stud Heal. Network, Knowledge-Based Syst. 121 (2017) 173–184.
Technol Inf. 229 (2016) 335–344. [29] T.L.M. Van Kasteren, G. Englebienne, B.J.A. Kröse, Activity [4] V.G. Sánchez, I. Taylor, P.C. Bing-Jonsson, Ethics of Smart House recognition using semi-markov models on real world smart home
Welfare Technology For Older Adults: A Systematic Literature datasets, J. Ambient Intell. Smart Environ. 2 (2010) 311–325.
Review, Int. J. Technol. Assess. Health Care. (2017) 1–9. [30] S. Ahmad, A. Lavin, S. Purdy, Z. Agha, Unsupervised real-time [5] V. Chandola, A. Banerjee, V. Kumar, Anomaly detection: A anomaly detection for streaming data, Neurocomputing. 262 (2017)
survey, ACM Comput. Surv. 41 (2009) 15. 134–147.
[6] M.R. Alam, M.B.I. Reaz, M. Ali, S.A. Samad, F.H. Hashim, M.K. [31] Y. Cui, S. Ahmad, J. Hawkins, Continuous online sequence Hamzah, Human activity classification for smart home: A learning with an unsupervised neural network model, Neural multiagent approach, in: Ind. Electron. Appl. (ISIEA), 2010 IEEE Comput. 28 (2016) 2474–2504.
Symp., 2010: pp. 511–514. [32] E.N. Osegi, Using the hierarchical temporal memory spatial pooler
[7] S.T.M. Bourobou, Y. Yoo, User Activity Recognition in Smart for short-term forecasting of electrical load time series, Appl.
Homes Using Pattern Clustering Applied to Temporal ANN Comput. Informatics. (2018).
Algorithm, Sensors. 15 (2015) 11953–11971. [33] T. V Duong, D.Q. Phung, H.H. Bui, S. Venkatesh, Human behavior [8] F. Jelinek, Interpolated estimation of Markov source parameters recognition with generic exponential family duration modeling in
from sparse data, in: Proc. Work. Pattern Recognit. Pract. 1980, the hidden semi-markov model, in: Pattern Recognition, 2006.
1980. ICPR 2006. 18th Int. Conf., 2006: pp. 202–207.
[9] L.R. Rabiner, A tutorial on hidden Markov models and selected [34] J. O’Connell, S. Højsgaard, others, Hidden semi markov models for
applications in speech recognition, Proc. IEEE. 77 (1989) 257–286. multiple observation sequences: The mhsmm package for R, J. Stat.
[10] T. V Duong, H.H. Bui, D.Q. Phung, S. Venkatesh, Activity Softw. 39 (2011) 1–22.
recognition and abnormality detection with the switching hidden [35] Ordonez, Activities of Daily Living (ADLs) Recognition Using semi-markov model, in: Comput. Vis. Pattern Recognition, 2005. Binary Sensors Data Set, (2013).
CVPR 2005. IEEE Comput. Soc. Conf., 2005: pp. 838–845. https://archive.ics.uci.edu/ml/datasets/Activities+of+Daily+Living+
[11] D. Bruckner, B. Sallans, R. Lang, Behavior learning via state chains 28ADLs29+Recognition+Using+Binary+Sensors.
from motion detector sensors, in: Bio-Inspired Model. Network, [36] F.J. Ordónez, P. de Toledo, A. Sanchis, Activity recognition using Inf. Comput. Syst. 2007. Bionetics 2007. 2nd, 2007: pp. 176–183. hybrid generative/discriminative models on home environments [12] T. Starner, J. Auxier, D. Ashbrook, M. Gandy, The gesture pendant: using binary sensors, Sensors. 13 (2013) 5460–5477.
A self-illuminating, wearable, infrared computer vision system for [37] S.-Z. Yu, Hidden semi-Markov models, Artif. Intell. 174 (2010) home automation control and medical monitoring, in: Wearable 215–243.
Comput. Fourth Int. Symp., 2000: pp. 87–94.
[13] C.A. Detweiler, K. V Hindriks, A survey of values, technologies and contexts in pervasive healthcare, Pervasive Mob. Comput. 27 (2016) 1–13.
[14] E.D. Mynatt, A.-S. Melenhorst, A.D. Fisk, W. Rogers, others, Aware technologies for aging in place: understanding user needs and attitudes, Pervasive Comput. IEEE. 3 (2004) 36–41.
[15] H. Zheng, H. Wang, N. Black, Human activity detection in smart home environment with self-adaptive neural networks, in:
